-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
489 lines (372 loc) · 21 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision.models import resnet50
from torch.utils.data import DataLoader
from torchvision import transforms
from accelerate import Accelerator
from tqdm.auto import tqdm
import os
from omegaconf import OmegaConf
from datasets import load_dataset
from model import IRFD, IRFDLoss, StyleGANLoss
from hsemotion_onnx.facial_emotions import HSEmotionRecognizer
from torchvision.utils import save_image
from CelebADataset import CelebADataset,ProgressiveDataset,AffectNetDataset
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torch.optim import Adam
from torch.utils.tensorboard import SummaryWriter
from torch.optim import AdamW
import random
import numpy as np
import torch.nn.functional as F
from PIL import Image, ImageDraw, ImageFont
import torchvision.transforms as transforms
import time
def save_debug_images(x_s, x_t, x_s_recon, x_t_recon, step, resolution, output_dir):
def denorm(x):
return (x * 0.5 + 0.5).clamp(0, 1)
x_s, x_t = denorm(x_s), denorm(x_t)
x_s_recon, x_t_recon = denorm(x_s_recon), denorm(x_t_recon)
combined = torch.cat([x_s, x_s_recon, x_t, x_t_recon], dim=0)
num_sets = min(16, x_s.size(0))
save_image(combined[:num_sets*4], os.path.join(output_dir, f"debug_step_{step}_resolution_{resolution}.png"), nrow=4)
# save with emotion labels
def save_emotion_debug_images(x_s, x_t, x_s_recon, x_t_recon, emotion_labels_s, emotion_labels_t, step, resolution, output_dir):
def denorm(x):
return (x * 0.5 + 0.5).clamp(0, 1)
def add_emotion_text(img_tensor, emotion_label):
img = transforms.ToPILImage()(img_tensor)
draw = ImageDraw.Draw(img)
font = ImageFont.load_default()
draw.text((5, 5), f"Emotion: {emotion_label}", (255, 255, 255), font=font)
return transforms.ToTensor()(img)
x_s, x_t = denorm(x_s), denorm(x_t)
x_s_recon, x_t_recon = denorm(x_s_recon), denorm(x_t_recon)
# Add emotion labels to images
x_s = torch.stack([add_emotion_text(img, label) for img, label in zip(x_s, emotion_labels_s)])
x_t = torch.stack([add_emotion_text(img, label) for img, label in zip(x_t, emotion_labels_t)])
x_s_recon = torch.stack([add_emotion_text(img, label) for img, label in zip(x_s_recon, emotion_labels_s)])
x_t_recon = torch.stack([add_emotion_text(img, label) for img, label in zip(x_t_recon, emotion_labels_t)])
combined = torch.cat([x_s, x_s_recon, x_t, x_t_recon], dim=0)
num_sets = min(16, x_s.size(0))
save_image(combined[:num_sets*4], os.path.join(output_dir, f"emotion_step_{step}_resolution_{resolution}.png"), nrow=4)
def create_progressive_dataloader(config, base_dataset, resolution, is_validation=False):
progressive_dataset = ProgressiveDataset(base_dataset, resolution)
# return torch.utils.data.DataLoader(
# OverfitDataset('S.png', 'T.png'),
# batch_size=1,
# num_workers=config.training.num_workers,
# pin_memory=True
# )
# Split the dataset into training and validation
train_size = int(0.8 * len(progressive_dataset)) # 80% for training
val_size = len(progressive_dataset) - train_size
train_dataset, val_dataset = torch.utils.data.random_split(progressive_dataset, [train_size, val_size])
if is_validation:
dataset = val_dataset
batch_size = config.training.eval_batch_size
shuffle = False
else:
dataset = train_dataset
batch_size = config.training.train_batch_size
shuffle = True
return torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
shuffle=shuffle,
num_workers=config.training.num_workers,
pin_memory=True,
persistent_workers=True,
prefetch_factor=2
)
def log_training_step(writer, loss, l_identity, l_cls, l_pose, l_emotion, l_self, global_step, resolution):
writer.add_scalar(f'Loss/Total/Resolution_{resolution}', loss.item(), global_step)
writer.add_scalar(f'Loss/Identity/Resolution_{resolution}', l_identity.item(), global_step)
writer.add_scalar(f'Loss/Classification/Resolution_{resolution}', l_cls.item(), global_step)
writer.add_scalar(f'Loss/Pose/Resolution_{resolution}', l_pose.item(), global_step)
writer.add_scalar(f'Loss/Emotion/Resolution_{resolution}', l_emotion.item(), global_step)
writer.add_scalar(f'Loss/SelfReconstruction/Resolution_{resolution}', l_self.item(), global_step)
def compute_gradient_penalty(discriminator, real_samples, fake_samples):
batch_size = real_samples.size(0)
alpha = torch.rand(batch_size, 1, 1, 1).to(real_samples.device)
interpolates = (alpha * real_samples + ((1 - alpha) * fake_samples)).requires_grad_(True)
d_interpolates = discriminator(interpolates)
fake = torch.ones(d_interpolates.size()).to(real_samples.device)
gradients = torch.autograd.grad(
outputs=d_interpolates,
inputs=interpolates,
grad_outputs=fake,
create_graph=True,
retain_graph=True,
only_inputs=True,
)[0]
# Use .reshape() instead of .view()
gradients = gradients.reshape(batch_size, -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
return gradient_penalty
def train_epoch(config, model, dataloader, optimizer_G, optimizer_D, criterion, accelerator, epoch, writer):
model.train()
total_loss_G = 0
total_loss_D = 0
progress_bar = tqdm(total=len(dataloader), disable=not accelerator.is_local_main_process)
progress_bar.set_description(f"Epoch {epoch+1}")
# Label smoothing
real_label = 0.9
fake_label = 0.1
def add_instance_noise(x, std=0.1):
return x + torch.randn_like(x) * std
for step, batch in enumerate(dataloader):
with accelerator.accumulate(model):
x_s, x_t = batch["source_image"], batch["target_image"]
emotion_labels_s, emotion_labels_t = batch["emotion_labels_s"], batch["emotion_labels_t"]
# Train Discriminator
optimizer_D.zero_grad()
# Real images
d_real_s = model.D(add_instance_noise(x_s))
d_real_t = model.D(add_instance_noise(x_t))
loss_D_real = (F.binary_cross_entropy_with_logits(d_real_s, torch.full_like(d_real_s, real_label)) +
F.binary_cross_entropy_with_logits(d_real_t, torch.full_like(d_real_t, real_label))) / 2
# Fake images
with torch.no_grad():
outputs = model(x_s, x_t)
x_s_recon, x_t_recon = outputs[0], outputs[1]
d_fake_s = model.D(add_instance_noise(x_s_recon.detach()))
d_fake_t = model.D(add_instance_noise(x_t_recon.detach()))
loss_D_fake = (F.binary_cross_entropy_with_logits(d_fake_s, torch.full_like(d_fake_s, fake_label)) +
F.binary_cross_entropy_with_logits(d_fake_t, torch.full_like(d_fake_t, fake_label))) / 2
# R1 regularization
r1_reg_s = compute_r1_reg(model.D, x_s)
r1_reg_t = compute_r1_reg(model.D, x_t)
r1_reg = (r1_reg_s + r1_reg_t) / 2
loss_D = loss_D_real + loss_D_fake + config.training.r1_weight * r1_reg
accelerator.backward(loss_D)
optimizer_D.step()
# Train Generator (less frequently)
if step % config.training.G_steps == 0:
optimizer_G.zero_grad()
outputs = model(x_s, x_t)
x_s_recon, x_t_recon, fi_s, fe_s, fp_s, fi_t, fe_t, fp_t, _, _ = outputs
l_pose_landmark, l_emotion, l_identity, l_recon = criterion(
x_s, x_t, x_s_recon, x_t_recon, fi_s, fe_s, fp_s, fi_t, fe_t, fp_t,
emotion_labels_s, emotion_labels_t
)
d_fake_s = model.D(x_s_recon)
d_fake_t = model.D(x_t_recon)
loss_G_adv = (F.binary_cross_entropy_with_logits(d_fake_s, torch.full_like(d_fake_s, real_label)) +
F.binary_cross_entropy_with_logits(d_fake_t, torch.full_like(d_fake_t, real_label))) / 2
loss_G = l_pose_landmark + l_emotion + l_identity + l_recon + config.training.stylegan_loss_weight * loss_G_adv
accelerator.backward(loss_G)
if config.training.grad_clip:
accelerator.clip_grad_norm_(model.parameters(), config.training.grad_clip_value)
optimizer_G.step()
if accelerator.sync_gradients:
progress_bar.update(1)
total_loss_G += loss_G.detach().float() if 'loss_G' in locals() else 0
total_loss_D += loss_D.detach().float()
global_step = epoch * len(dataloader) + step
if accelerator.is_main_process:
if step % config.training.logging_steps == 0:
print(f"Epoch {epoch+1}, Step {step}: loss_G = {loss_G.item() if 'loss_G' in locals() else 0:.4f}, loss_D = {loss_D.item():.4f}")
writer.add_scalar('Loss/Train/Generator', loss_G.item() if 'loss_G' in locals() else 0, global_step)
writer.add_scalar('Loss/Train/Discriminator', loss_D.item(), global_step)
writer.add_scalar('Loss/Train/Pose_Landmark', l_pose_landmark.item() if 'l_pose_landmark' in locals() else 0, global_step)
writer.add_scalar('Loss/Train/Emotion', l_emotion.item() if 'l_emotion' in locals() else 0, global_step)
writer.add_scalar('Loss/Train/Identity', l_identity.item() if 'l_identity' in locals() else 0, global_step)
writer.add_scalar('Loss/Train/Reconstruction', l_recon.item() if 'l_recon' in locals() else 0, global_step)
writer.add_scalar('Loss/Train/R1_Regularization', r1_reg.item(), global_step)
if global_step % config.training.save_image_steps == 0:
save_debug_images(x_s, x_t, x_s_recon, x_t_recon, epoch, step, config.training.output_dir)
if global_step % config.training.save_steps == 0:
save_path = os.path.join(config.training.output_dir, f"best_model-epoch-{epoch+1}-{global_step}")
accelerator.save({
'model_state_dict': model.state_dict(),
'optimizer_G': optimizer_G.state_dict(),
'optimizer_D': optimizer_D.state_dict(),
'epoch': epoch,
'resolution': model.current_resolution,
'config': config,
}, save_path)
return total_loss_G.item() / len(dataloader), total_loss_D.item() / len(dataloader)
def compute_r1_reg(D, real_img):
real_img = real_img.requires_grad_(True)
real_pred = D(real_img)
grad_real = torch.autograd.grad(
outputs=real_pred.sum(), inputs=real_img, create_graph=True
)[0]
grad_penalty = grad_real.pow(2).reshape(grad_real.shape[0], -1).sum(1).mean()
return grad_penalty
def compute_path_length_regularization(G, latents, gen_img1, gen_img2):
noise = torch.randn_like(gen_img1) / (gen_img1.shape[2] * gen_img1.shape[3]) ** 0.5
grad1, grad2 = torch.autograd.grad(
outputs=(gen_img1 * noise).sum() + (gen_img2 * noise).sum(),
inputs=latents,
create_graph=True,
only_inputs=True,
)[0]
path_lengths = torch.sqrt(grad1.pow(2).sum(2).mean(1) + grad2.pow(2).sum(2).mean(1))
pl_mean = path_lengths.mean()
pl_length = ((path_lengths - pl_mean) ** 2).mean()
return pl_length
def validate(config, model, dataloader, criterion, accelerator, stylegan_loss, current_resolution):
model.eval()
total_loss = 0
with torch.no_grad():
for batch in dataloader:
x_s, x_t = batch["source_image"], batch["target_image"]
emotion_labels_s, emotion_labels_t = batch["emotion_labels_s"], batch["emotion_labels_t"]
print(f"Input shapes: x_s: {x_s.shape}, x_t: {x_t.shape}")
outputs = model(x_s, x_t)
x_s_recon, x_t_recon, fi_s, fe_s, fp_s, fi_t, fe_t, fp_t, _, _ = outputs
print(f"Reconstruction shapes: x_s_recon: {x_s_recon.shape}, x_t_recon: {x_t_recon.shape}")
print(f"Feature shapes: fi_s: {fi_s.shape}, fe_s: {fe_s.shape}, fp_s: {fp_s.shape}")
irfd_loss = criterion(
x_s, x_t, x_s_recon, x_t_recon,
fi_s, fe_s, fp_s, fi_t, fe_t, fp_t,
emotion_labels_s, emotion_labels_t
)
# Flatten and concatenate all features
fi_s_flat = fi_s.view(x_s.size(0), -1)
fe_s_flat = fe_s.view(x_s.size(0), -1)
fp_s_flat = fp_s.view(x_s.size(0), -1)
fi_t_flat = fi_t.view(x_t.size(0), -1)
fe_t_flat = fe_t.view(x_t.size(0), -1)
fp_t_flat = fp_t.view(x_t.size(0), -1)
combined_s = torch.cat([fi_s_flat, fe_s_flat, fp_s_flat], dim=1)
combined_t = torch.cat([fi_t_flat, fe_t_flat, fp_t_flat], dim=1)
print(f"Combined feature shapes: combined_s: {combined_s.shape}, combined_t: {combined_t.shape}")
try:
fake_s = model.Gd(combined_s, current_resolution)
fake_t = model.Gd(combined_t, current_resolution)
print(f"Generated fake shapes: fake_s: {fake_s.shape}, fake_t: {fake_t.shape}")
except RuntimeError as e:
print(f"Error in Gd forward pass: {str(e)}")
print(f"Gd input_dim: {model.Gd.input_dim}")
raise e
stylegan_loss_value = stylegan_loss(x_s, fake_s) + stylegan_loss(x_t, fake_t)
loss = sum(irfd_loss) + config.training.label_balance * stylegan_loss_value
total_loss += loss.detach().float()
return total_loss.item() / len(dataloader)
def main():
config = OmegaConf.load("config.yaml")
accelerator = Accelerator(
mixed_precision=config.training.mixed_precision,
gradient_accumulation_steps=config.training.gradient_accumulation_steps,
log_with=config.logging.log_with,
project_dir=os.path.join(config.training.output_dir, "logs"),
)
if accelerator.is_main_process:
os.makedirs(config.training.output_dir, exist_ok=True)
OmegaConf.save(config, os.path.join(config.training.output_dir, "config.yaml"))
model = IRFD(max_resolution=256)
optimizer_G = torch.optim.Adam(model.Gd.parameters(), lr=config.training.G_lr)
optimizer_D = torch.optim.Adam(model.D.parameters(), lr=config.training.D_lr)
criterion = IRFDLoss(config,accelerator.device)
# Check for existing checkpoint
start_epoch = 0
latest_checkpoint = None
if os.path.exists(config.training.output_dir):
checkpoints = [f for f in os.listdir(config.training.output_dir) if f.startswith("best_model")]
if checkpoints:
latest_checkpoint = max(checkpoints, key=lambda x: os.path.getmtime(os.path.join(config.training.output_dir, x)))
latest_checkpoint = os.path.join(config.training.output_dir, latest_checkpoint)
print(f"👑 Found latest checkpoint: {latest_checkpoint}")
# Load checkpoint if exists
if latest_checkpoint:
checkpoint = torch.load(latest_checkpoint, map_location=accelerator.device)
model.load_state_dict(checkpoint['model_state_dict']) # Load the state dict
optimizer_G.load_state_dict(checkpoint['optimizer_G'])
optimizer_D.load_state_dict(checkpoint['optimizer_D'])
start_epoch = checkpoint['epoch'] + 1
current_resolution = checkpoint['resolution']
config = checkpoint['config']
print(f"Resumed training from epoch {start_epoch}, resolution {current_resolution}")
# Set up preprocessing
preprocess = transforms.Compose([
transforms.Resize((256, 256)), # Start with the highest resolution
# transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
])
def get_base_dataset(preprocess):
test = AffectNetDataset(
root_dir="/media/oem/12TB/AffectNet/train",
preprocess=preprocess,
remove_background=False,
use_greenscreen=False,
cache_dir='/media/oem/12TB/AffectNet/train/cache'
)
return test
# return CelebADataset(config.dataset.name, config.dataset.split, preprocess)
# Load the dataset
base_dataset = get_base_dataset(preprocess)
train_dataloader = create_progressive_dataloader(config, base_dataset, 64, is_validation=False)
val_dataloader = create_progressive_dataloader(config, base_dataset, 64, is_validation=True)
model, optimizer_G, optimizer_D, train_dataloader, val_dataloader, criterion = accelerator.prepare(
model, optimizer_G, optimizer_D, train_dataloader, val_dataloader, criterion
)
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
stylegan_loss = StyleGANLoss(accelerator.device)
scheduler_G = ReduceLROnPlateau(optimizer_G, mode='min', factor=0.1, patience=config.training.early_stopping_patience)
scheduler_D = ReduceLROnPlateau(optimizer_D, mode='min', factor=0.1, patience=config.training.early_stopping_patience)
writer = SummaryWriter(log_dir=os.path.join(config.training.output_dir, "logs"))
resolutions = [256] #64, 128,
epochs_per_resolution = config.training.epochs_per_resolution
for resolution in resolutions:
print(f"Training at resolution: {resolution}x{resolution}")
model.adjust_for_resolution(resolutions[resolutions.index(resolution)])
preprocess = transforms.Compose([
transforms.Resize((resolution, resolution)),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
])
# preprocess = transforms.Compose([
# transforms.Resize((resolution, resolution)),
# transforms.RandomHorizontalFlip(),
# transforms.RandomAffine(degrees=10, translate=(0.1, 0.1), scale=(0.9, 1.1), shear=10),
# transforms.RandomPerspective(distortion_scale=0.2, p=0.5),
# transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
# transforms.RandomApply([transforms.ElasticTransform(alpha=250.0, sigma=8.0)], p=0.5),
# transforms.ToTensor(),
# transforms.Normalize([0.5], [0.5])
# ])
base_dataset = get_base_dataset(preprocess)
train_dataloader = create_progressive_dataloader(config, base_dataset, resolution, is_validation=False)
val_dataloader = create_progressive_dataloader(config, base_dataset, resolution, is_validation=True)
model, optimizer_G, optimizer_D, train_dataloader, val_dataloader, criterion = accelerator.prepare(
model, optimizer_G, optimizer_D, train_dataloader, val_dataloader, criterion
)
scheduler_G = ReduceLROnPlateau(optimizer_G, mode='min', factor=0.1, patience=config.training.early_stopping_patience)
scheduler_D = ReduceLROnPlateau(optimizer_D, mode='min', factor=0.1, patience=config.training.early_stopping_patience)
best_val_loss = float('inf')
for epoch in range(epochs_per_resolution):
train_loss_G, train_loss_D = train_epoch(config, model, train_dataloader, optimizer_G, optimizer_D, criterion, accelerator, epoch, writer)
val_loss = validate(config, model, val_dataloader, criterion, accelerator, stylegan_loss, resolution)
if accelerator.is_main_process:
print(f"Resolution: {resolution}, Epoch {epoch+1}/{epochs_per_resolution}: train_loss_G = {train_loss_G:.4f}, train_loss_D = {train_loss_D:.4f}, val_loss = {val_loss:.4f}")
writer.add_scalar(f'Loss/Train/Generator/Resolution_{resolution}', train_loss_G, epoch)
writer.add_scalar(f'Loss/Train/Discriminator/Resolution_{resolution}', train_loss_D, epoch)
writer.add_scalar(f'Loss/Validation/Resolution_{resolution}', val_loss, epoch)
# Save debug images
with torch.no_grad():
x_s, x_t = next(iter(val_dataloader))["source_image"], next(iter(val_dataloader))["target_image"]
outputs = model(x_s, x_t)
x_s_recon, x_t_recon = outputs[0], outputs[1]
save_debug_images(x_s, x_t, x_s_recon, x_t_recon, epoch, resolution, config.training.output_dir)
scheduler_G.step(val_loss)
scheduler_D.step(val_loss)
if val_loss < best_val_loss:
best_val_loss = val_loss
if accelerator.is_main_process:
accelerator.save({
'model': model,
'optimizer_G': optimizer_G.state_dict(),
'optimizer_D': optimizer_D.state_dict(),
'epoch': epoch,
'resolution': resolution,
'config': config,
}, os.path.join(config.training.output_dir, f"checkpoint-resolution-{resolution}-epoch-{epoch}.pth"))
accelerator.end_training()
writer.close()
if __name__ == "__main__":
main()